8  Zooarchaeology

Page under construction

8.1 Case studies

The following map shows the sites under investigation, divided by chronology. Please select the desired chronology (or chronologies) from the legend on the right. Legend: R = Roman, LR = Late Roman, EMA = Early Middle Ages, Ma = 11th c. onwards

The faunal dataset used in this study is both extensive and diverse, containing over 466 records. While NISP is a useful proxy for historical animal farming, it is not without its limitations, as previously discussed in the methods section. What is crucial to emphasize here, however, is the presence of overdispersion within the data, which necessitates more sophisticated and nuanced approaches than simply calculating overall means for each animal species. Overdispersion is a common occurrence when analyzing datasets of this nature, given the unique factors at play in each context, including historical and depositional influences. Nevertheless, data modeling requires simplification and causal reasoning, and the best approach is to start with straightforward models that can account for overdispersion and generate credible distributions. To this end, Bayesian hierarchical models were developed for each chronology, context type, macroregion, and geography. As a further step, an additional analysis was conducted that solely focused on altitude and chronology as predictors. By examining the specific contributions of altitude and chronology in predicting animal farming/consumption patterns, this analysis provides valuable insights that complement the earlier models. These findings underscore the importance of considering multiple factors when studying the probability of occurrence of farmed and wild animals in historical contexts. They also demonstrate the potential benefits of simplified models to focus on key predictors. Moreover, several attempts were made to create a coherent understanding of the likelihood of economically valuable animals occurring during the first millennium, in order to provide a more comprehensive perspective on the role of animal farming in shaping historical societies.

8.2 Data exploration

As previously mentioned, the faunal dataset used in this study exhibits overdispersion, which is a common issue when analysing datasets of this type. The term ‘overdispersion’ refers to the presence of greater variability in the data than expected based on a normal curve (Figure 8.1). In this section, I will present the distribution of the animals of interest to provide a visual representation of the dispersion within the dataset. By examining the distribution curves, we can better understand the variability that exist within the dataset, and can use this information to develop more accurate models. In Figure 8.2, it is evident that the distribution of animal remains in the faunal dataset is not symmetrical and does not conform to a normal curve. This non-normal distribution indicates that the standard measures of central tendency, such as mean and median, may not accurately capture the overall distribution of the data. Traditional statistical measures, such as measures of central tendency and dispersion, are commonly used in frequentist approaches to analyse data. However, these measures may not be appropriate for analysing the complex patterns of animal farming and consumption in the faunal dataset due to the presence of overdispersion. As an alternative, Bayesian multilevel models can account for overdispersion by incorporating appropriate probability distributions, such as the betabinomial distribution.

Figure 8.1: Probability density curves of three simulated normal distributions, representing underdispersed (sd = 0.5), normal (sd = 1), and overdispersed (sd = 2) data.

Figure 8.2: Distribution of animal remains in the dataset, displayed both by NISP raw value and NISP proportions.

8.3 Chronology

8.4 Context type

Chronological models provide a broad overview of the trends in domestic and wild animal husbandry and consumption. However, it is important to consider the context where the animal remains are found because different site types can influence the NISP of a particular animal. We can stratify the dataset by site type, but there is a challenge: the four chronologies we are studying also affect the availability, frequency, and economic dynamics of the site types. For example, the agricultural strategies employed by rural villas during the Roman and Late Roman periods could differ based on political and economic events, which could affect the relative amount of animal NISP. To better understand how the context where animal remains were found affects our outcome variable, the NISP, we can use a Directed Acyclic Graph (DAG). In this DAG, the Chronology node is a parent of the Site_Type node, and it affects the NISP count directly and indirectly through Site_Type. To account for this backdoor path, we block by Chronology and stratify the dataset by Site_Type. Since we have two categorical predictors of interest, we can use an interaction index variable to generate distinct intercepts for each chronology and context type. For instance, the Roman:Rural context type could be assigned an index of 1, while Roman:Urban could be assigned an index of 2, and so on. This interaction dummy index (\({[TCid]}\)) will determine the variation in intercepts (\(\alpha\)) across different context types.

Figure 8.4: This is a Directed Acyclic Graph (DAG) that represents the relationship between Chronology, Site_Type, and NISP in the study. Chronology is shown as a collider variable, and its influence on NISP is mediated through Site_Type. To block the backdoor path between Chronology and NISP, the dataset is stratified by Chronology.

The approach for the proposed model for estimating the posterior likelihood of animal occurrence in various chronological phases and context types is very to the models discussed earlier. The only modification is the interaction index. The model structure and priors remain the same as in the previous models. There are several contexts that have been recorded in the database (see Table 5.1 for a full list), but the model required some simplifications in order to avoid extremely small categories. For instance, religious monasteries have been grouped with other Religious sites, or Roman mansiones have been classified as Rural for these models. Categorical simplifications can obscure the nuances of data, but tiny sample sizes produce very uninformative posterior distributions.

\[ A_{i} \sim BetaBinomial(NISP_{i}, \bar{p}_{i} , \phi_{i}) \]

\[ logit(\bar{p}_{i}) = \alpha_{[TCid]} \]

\[ \alpha_{[TCid]} \sim Normal(0,1.5) \]

\[ \phi_{[TCid]} \sim Exponential(1) + 2 \]

8.4.1 Pigs

As previously noted, pig remains are consistently the most frequently found type of faunal remains in first millennium Italian excavations. However, this study reveals divergent patterns in the presence of pigs depending on the site type and function. We will first discuss categories with narrower credible intervals before moving on to those with smaller sample sizes and wider credible intervals. Urban sites are the most common category, where the probability of finding pig remains follows patterns similar to the chronological trends presented earlier. During the Roman phase, the probability of finding pig remains in urban contexts is very high compared to other animals or categories, with a mean probability of almost 0.5. The consumption of pork in cities appears to have decreased during the Late Roman and Early Medieval phase, with similar credible intervals, only to increase again in the 11th century. Possible hypotheses will be discussed later on. Although the precision of the pig beta posterior in urban contexts is slightly below the across-contexts mean of 5, the curves are not too dispersed, at least in the Roman and Late Roman periods, and we can be confident in these results. On the other hand, the 95% HDIs for pigs in rural contexts are lower than those in urban contexts during the Roman and Late Roman periods. In the Early Medieval phase, the mean probabilities are comparable to urban contexts, with both categories having a mean of around 0.31, which then increases to 0.37 in the 11th century (0.39 in urban contexts). In general, while the decline in pig presence appears to be significant in urban areas, rural contexts exhibit only a slight increase. Rural villas have been analyzed separately from other rural sites due to their potential implications for the production and consumption of meat, given their status as elite sites. Although the sample size is not large enough to provide highly confident posterior predictions, the results suggest that pork was likely an important component of the diet in these sites, with a mean of approximately 0.4 in the Roman period and an increase in the Late Roman phase. However, the Early Medieval phase is characterized by greater uncertainty, as the sample size is reduced and rural estates often changed function or were abandoned. The credible interval for this period ranges from 0.25 to 0.62. The suggested increase in pigs NISP in the Late Roman period will be discussed further later, as the historical debate might provide helpful hypotheses. It is worth noting that there is no sample from the 11th century for rural villas, which is why that line is not included in the graph. Fortified sites were also large consumers of pigs, with the probabilities of occurrence sticking around the mean value and a minor increase in the Early Medieval-Medieval periods. However, there are no significant changes in trends to observe. The probabilities of occurrence in the Late Roman phase are not so trustworthy as the credible interval is large. It is worth noting that there is no Roman posterior prediction on the graph because there were not fortified Roman sites in the sample except for one, the castrum of Ostia (MacKinnon, 2014), which has not been included in the observed data for this model as it was one single sample. The castrum had a 71.9% NISP proportion of pigs, out of a sample of 121 total NISP. The religious category has fewer observations than the site types presented before, and the credible intervals are wider. The Roman and Late Roman religious sites mostly include temples such as the temples C and D of Grumentum, the Demetra temple in Macchia delle Valli, the Mithraeum of Crypta Balbi, and more. In the Medieval phases, religious sites are mostly monasteries including the well-studied ones of Monte Gelato, Farfa, San Salvo, S. Giulia of Brescia, and San Vincenzo al Volturno. The range of credible intervals can span probabilities of 0.40, so we must be cautious in our considerations. If we look at the mean of the HDIs, a precise pattern does not emerge. However, in every chronology, the probability of occurrence of pig’s NISP is over the across-contexts average of 0.41. The credible intervals for necropolis sites are too large - spanning almost the entire probability range - to trace any clear patterns. With only nine samples from seven sites, including the Roman and Late Roman necropoleis of Cantone, Otranto, San Cassiano (Riva del Garda), Trieste (loc. Crosada), Poggio Gramignano, San Lorenzo di Sebato, and the Early Medieval necropolis of Baggiovara, the sample size is small. As a result, it is best to draw only qualitative conclusions in the later discussion. Finally, the last category includes only one well-known site, the Flavium amphitheater in Rome, which provided 29 zooarchaeological samples. The 95% HDIs for this site show very high probabilities, with mean values of 0.65 for the Roman and Late Roman periods and 0.49 for the Early Medieval period. However, the credible interval is quite wide for the Early Medieval phase, making it less certain. This type of site was not grouped with the other urban observations as it was too unique.

8.4.2 Cattle

8.4.3 Caprine

8.4.4 Edible W. Mammals

8.4.5 Community plot

8.5 Macroregion

To estimate the animals’ occurrences probability in each chronology and macroregion, I used a betabinomial distribution to model overdispersion in the data. The \(A\) on the left side of the formula is the outcome variable—the animal NISP counts for each observation \(i\). This is an intercept-only model, where the intercept \(\alpha\) carries an interaction index \({[REGid]}\) as the model will provide estimates for each macroregion and chronology under examination. The \(\phi\) parameter indicates the precision in the Beta distribution, modelled by chronology and macroregion.

\[ A_{i} \sim BetaBinomial(NISP_{i}, \bar{p}_{i} , \phi_{i}) \]

\[ logit(\bar{p}_{i}) = \alpha_{[REGid]} \]

\[ \alpha_{[REGid]} \sim Normal(0,1.5) \]

\[ \phi_{[REGid]} \sim Exponential(1)+2 \]

8.5.1 Pigs

8.5.2 Cattle

8.5.3 Caprine

8.5.4 Edible W. Mammals

8.6 Geography

Animals distributions can vary across different geographical features. This research has considered plain, coast, hill and mountains as the most common geographical features in the Italian peninsula. Archaeological excavations where zooarchaeological remains have been analysed are located at low altitudes. Although as expected there are more mountain sites in Northern Italy, sampled sites are evenly placed on plains, coastlands and hills across the three Italian macroregions.

Figure 8.5: Distribution of sites on different geographical features.

To estimate the animals’ occurrences probability in each chronology and geography, I used a betabinomial distribution to model overdispersion in the data. The \(A\) on the left side of the formula is the outcome variable—the animal NISP counts for each observation \(i\). This is an intercept-only model, where the intercept \(\alpha\) carries an interaction index \({[GEOid]}\) as the model will provide estimates for each geography and chronology under examination. The \(\phi\) parameter indicates the precision in the Beta distribution, modelled by chronology and geography.

\[ A_{i} \sim BetaBinomial(NISP_{i}, \bar{p}_{i} , \phi_{i}) \]

\[ logit(\bar{p}_{i}) = \alpha_{[GEOid]} \]

\[ \alpha_{[GEOid]} \sim Normal(0,1.5) \]

\[ \phi_{[GEOid]} \sim Exponential(1)+2 \]

8.6.1 Pigs

8.6.2 Cattle

8.6.3 Caprine

8.6.4 Edible W. Mammals

8.7 Altitude

The probability of occurrence of the most common faunal remains can be modelled against the elevation of sites in the four time periods under consideration. It is worth noting that the sites where the zooarchaeological remains have been found are not evenly distributed. In the Roman age, most sites investigated are located between 0 and 100 MSL, whereas after there is an increasing number of remains from sites between 100 and 400 MSL. Whether this reflects a real shift in settlement patterns is outside the aims of this study, but it might still be informative to visualise the different distribution of sites across elevations.

The proposed model to estimate the probability of occurrence as related to the altitude (the slope \(\beta\)) and chronology (\({[ChrID]}\)) uses a betabinomial distribution to model overdispersion in the data. The \(A\) on the left side of the formula is the outcome variable—the animal NISP counts for each observation \(i\). This is a simple intercept with slope model, where the intercept \(\alpha\) carries an index \({[ChrID]}\) as the model provides estimates for each chronology under examination. A single \(\phi\) parameter indicates the precision in the Beta distribution.

\[ A_{i} \sim BetaBinomial(NISP_{i}, \bar{p}_{i} , \phi_{i}) \]

\[ logit(\bar{p}_{i}) = \alpha_{[ChrID]} + \beta_{[ChrID]}\cdot Alt_{i} \]

\[ \alpha_{ChrID} \sim Normal(0,1.5) \]

\[ \beta_{ChrID} \sim Normal(0,1.5) \]

\[ \phi \sim Exponential(1)+2 \]

Figure 8.6: Prior predictive simulation for the altitude models used in this section.

8.7.1 Pigs

8.7.2 Cattle

8.7.3 Caprine

8.7.4 Edible W. Mammals

8.7.5 Community plot

Figure 8.7: MCMC estimates for slope and intercept plotted in the logit scale. Negative slopes indicate a negative relationship between the animal remains and increasing altitude. Intercepts were kept as a baseline occurrence probability of the species. Species on the left of the graph are rarer, species on the right are more common. It is important to notice that this represents the species response to elevation.